AI

Checking AI answers before they go out

3 min read

Dark green graphic reading “Erst prüfen, dann senden” with a list of three steps: opening the source and verifying the figure ticked, sending blindly not
Quick answer

You check AI answers with three moves. First, open the cited source yourself, because language models also invent references. Second, ask the counter-question: what speaks against this answer? Third, never adopt figures unchecked; every figure that leaves the business needs a primary source. Confirmation by the model itself checks nothing. The three moves cost minutes and catch the frequent and expensive mistakes, though not all of them.

This lesson is station 5 of the AI basics course and the third of three lessons; two tools sit in between. It covers the move that decides the value of all the others: checking before anything leaves the business.

Why is common sense not enough?

Because the usual warning signal is missing. With people, uncertainty shows in the tone. A language model can sound certain and be wrong – its tone tells you nothing. Invention and fact arrive in the same confident packaging. Whoever checks by feel is checking the packaging.

Move one: open the source yourself

Language models also invent source references, formally correct-looking ones with plausible titles. A reference is therefore not a check but an invitation to one. Open it, see whether the source exists and whether the statement is really in it.

The order matters: read the source first, then use the claim. The other way round, the invented figure travels into the proposal and the check never happens.

Move two: ask the counter-question

Ask the model for the opposing view: “What speaks against this answer? What would have to be true for it to be wrong?” That is not a real check, but it is a good error detector. Strikingly often the counter-question names exactly the weak spot.

What checks nothing, by contrast: asking the model to confirm its own answer. It does so far too often, even when the answer is wrong.

This lesson in two sentences:

You check AI answers with three moves: open the cited source yourself, ask the counter-question, and never adopt figures unchecked. Confirmation by the model itself is not a check, and the convincing tone of an answer says nothing about whether it is right.

Move three: never adopt figures unchecked

The strictest rule applies to figures, because they err at the highest cost: prices, quantities, deadlines, legal provisions. Every figure that leaves the business has a primary source, or it stays out.

One limit we state openly: these three moves do not find every mistake. They find the frequent and the expensive ones, and together they cost a few minutes. More certainty only comes with expertise in the field concerned.

Exercise: fall for it once, on purpose

Ask the tool for a statistic about your industry, including a source. Switch off the tool's web search for this if it has one – with web search you are testing the web, not the model. Open the source. Check both: does it exist, and is the figure in it?

If both hold, you chose well. If not, you caught an invented source before it ended up in a quote. After that you are ready for the knowledge check, the final station of the course.